The artificial intelligence revolution has a dirty secret: it is ruinously expensive to run.
Every time you fire off a prompt to ChatGPT, every time a customer service bot answers a question, every time an AI model generates an image or writes a line of code, someone is burning compute. Lots of it. Training large language models gets the headlines — billions of dollars, tens of thousands of GPUs, enough electricity to power small nations — but the real money, the real ongoing cost, is inference. That is the act of actually running the model, processing your request, and spitting out a response. Inference is where the meters are spinning, and for many companies, those meters are spinning terrifyingly fast.
Enter Doubleword, an 11-person startup headquartered in London that is building what its founder calls "sovereign compute." Founded by Meryem Arik, a former consultant turned entrepreneur, Doubleword is tackling the inference problem head-on. Its pitch is simple, audacious, and increasingly resonant in a world nervous about AI dependency: make running AI models dramatically cheaper, and do it in a way that does not hand the keys to American technology giants.
In an interview with Sifted at a sunlit café on Britton Street, Arik laid out the scale of the opportunity. "The compute market is massive," she said. "And the current pricing is essentially a tax on doing business with AI." Doubleword believes it can cut those costs by a factor of one hundred — not over decades, but every year. It is a staggering claim, one that would sound like the usual startup bluster if not for the growing roster of European clients signing on and the caliber of investors quietly backing the company.
The Inference Tax Nobody Talks About
Here is how the AI economy works in 2026. Companies like OpenAI, Anthropic, and Google have spent billions training foundation models — GPT-5.5, Claude Opus 4, Gemini Ultra. These models are powerful, general-purpose, and, crucially, hosted. If you want to use them, you pay per token, per request, per month. The pricing is opaque, the margins are enormous, and the infrastructure is controlled by a handful of American corporations.
For startups and enterprises alike, this creates a bind. AI is no longer optional — it is the baseline for competing. But the cost of integrating AI into products, workflows, and customer experiences scales alarmingly with usage. A company that ships an AI-powered feature to a million users can find its cloud bill ballooning overnight. The more successful you are, the more you pay. It is the opposite of the traditional software economics where marginal costs approach zero.
Inference, in other words, is the chokepoint. Training is a one-time (albeit massive) expense. Inference is forever. And as AI models get larger, smarter, and more deeply embedded in every application, the inference bill only grows. Analysts at Bernstein Research estimate that global AI inference spending will exceed training spending by 2027, and the gap will only widen from there.
"Everyone is focused on the next big model release," Arik told Sifted. "But the real question is: can you afford to actually use it at scale? Most companies can't. Not at today's prices."
Sovereign Compute: The Political Angle
Doubleword's solution is not just about cost. It is about control.
The concept of "sovereign compute" has been gaining traction across Europe, driven by a growing unease with American technological dominance. European regulators have already fined American tech giants billions for antitrust violations and data privacy breaches. European governments have launched initiatives to build homegrown AI capabilities — France's Mistral AI, Germany's Aleph Alpha, the European Union's own investments in supercomputing infrastructure. But the reality on the ground is that most European companies still route their AI workloads through AWS, Azure, Google Cloud, or directly to OpenAI's API servers in the United States.
This dependency has geopolitical implications. In a world of rising trade tensions, export controls on AI chips, and increasingly aggressive digital sovereignty policies, relying on foreign infrastructure for core business functions looks less like a convenience and more like a vulnerability. The Trump administration's previous term saw expanded restrictions on AI technology exports to China; a future administration could just as easily tighten rules on European access to American AI services, or demand data sharing as a condition of use.
Doubleword is positioning itself as the technical answer to this political anxiety. Its platform optimizes inference workloads so they can run efficiently on a wider range of hardware — not just the latest NVIDIA H100 clusters, but older GPUs, specialized inference chips, and eventually European-designed silicon. The goal is to let any company, anywhere, run powerful AI models on infrastructure they control, at a fraction of the cost of calling an American API.
"We are not asking companies to compromise on model quality," Arik emphasized. "We are asking them to question whether they need to pay a 10x premium for the privilege of having their data processed in Virginia."
How Doubleword Actually Works
The technical details of Doubleword's platform remain closely guarded, as befits a company in a competitive and technically complex space. But the broad strokes are discernible from Arik's public statements and the company's hiring patterns.
At its core, Doubleword appears to be building a model optimization and inference orchestration layer. Think of it as a highly intelligent traffic controller for AI workloads. Instead of sending every request to the most powerful (and expensive) model, Doubleword's system likely analyzes the complexity of each task and routes it to the most efficient model, hardware configuration, and execution path. Simple queries might run on a small, fine-tuned model on a single GPU. Complex reasoning tasks might be distributed across multiple optimized instances. The system learns from patterns, predicts load, and continuously squeezes efficiency out of the stack.
This approach mirrors a broader trend in the AI industry toward inference efficiency. Companies like Groq have built specialized inference chips that process tokens at mind-bending speeds. Startups like Fireworks AI and Together AI offer optimized inference APIs that undercut OpenAI's pricing. Open-source projects like vLLM and TensorRT-LLM squeeze more performance out of standard GPUs. Doubleword appears to be taking this further by adding a layer of model selection, workload prediction, and hardware abstraction that lets companies mix and match the most cost-effective backends.
The "sovereign" element comes from the ability to run this stack on any infrastructure — European clouds, on-premise servers, even edge devices — without being locked into a specific vendor's ecosystem. In Arik's vision, a German automotive supplier could run a fine-tuned quality-control model on local servers, a French bank could process sensitive documents on European sovereign cloud infrastructure, and a British startup could scale AI features without watching its burn rate explode.
The Market Opportunity
The numbers, if Doubleword delivers on even a fraction of its promises, are enormous.
The global market for AI inference is projected to grow from roughly $20 billion in 2025 to over $200 billion by 2030, according to multiple industry estimates. That growth is driven by the simple fact that every AI application that gets deployed needs to be run somewhere, somehow, and someone has to pay for the compute. Today, the lion's share of that spending flows to NVIDIA (for GPUs), cloud hyperscalers (for the infrastructure), and API providers like OpenAI (for the model access).
Doubleword is attempting to insert itself into that value chain as an efficiency layer — taking a cut of the savings it generates for customers. If it can truly deliver 10x or 100x cost reductions, its take rate could be substantial while still leaving customers far better off than the status quo. The business model is reminiscent of cloud cost optimization tools like CloudZero or Spot.io, but applied specifically to the AI inference stack.
The European angle is critical here. European companies spent an estimated €47 billion on cloud services in 2025, a significant portion of which went to American providers. European governments and regulators are actively pushing for "digital sovereignty" — the idea that critical data and infrastructure should be controlled within European borders. The EU's AI Act, while focused primarily on safety and transparency, also creates compliance burdens that make sovereign infrastructure more attractive. A European company storing customer data in an American cloud and processing it through an American AI model faces regulatory scrutiny that a fully European stack might avoid.
Doubleword's timing, in other words, is impeccable. The company is surfing a wave of technological necessity (inference costs must come down), political urgency (Europe wants tech independence), and regulatory pressure (the AI Act and data localization rules).
The Competitive Landscape
Doubleword is not alone in chasing the inference optimization prize. The space is increasingly crowded, and the incumbents are formidable.
On one side are the specialized inference providers — companies like Groq, SambaNova, Cerebras, and Tenstorrent that build custom hardware designed specifically for running AI models fast and cheap. These companies have raised billions in collective funding and are already landing major enterprise deals. Groq, for instance, has demonstrated inference speeds that make standard GPU clusters look sluggish, and its LPU (Language Processing Unit) architecture is purpose-built for the transformer models that dominate today's AI landscape.
On another side are the API optimization layers — Fireworks AI, Together AI, and Anyscale offer managed inference services that claim significant cost savings over calling OpenAI or Anthropic directly. These companies have raised hundreds of millions and have strong technical teams with deep roots in distributed systems and model serving.
Then there are the cloud hyperscalers themselves — AWS, Google Cloud, and Azure are all building their own inference optimization tools, from custom silicon (AWS's Inferentia and Trainium chips, Google's TPUs) to software stacks that promise better performance on standard hardware. These giants have scale, capital, and existing customer relationships that no startup can match.
Doubleword's differentiation, in Arik's telling, is the combination of deep optimization with true infrastructure independence. "We are not selling you another API that locks you in," she said. "We are giving you the tools to run AI anywhere, on anything, at a fraction of the cost." It is a compelling pitch, but one that will be tested against the reality of building and maintaining a platform that works across dozens of hardware configurations and model types.
The Founder and the Team
Meryem Arik is not a typical Silicon Valley AI founder. She does not have a PhD in machine learning from Stanford or a stint at Google Brain on her CV. Her background is in consulting — she worked at a major firm before striking out on her own — which gives her a grounding in the practical realities of enterprise technology adoption that many research-oriented founders lack.
What she lacks in academic credentials, she appears to make up for in clarity of vision and commercial hustle. Building an 11-person team that has already attracted investor interest and paying customers in one of the most competitive spaces in technology is no small feat. The team, while small, has been assembled with a clear focus on engineering talent — job postings suggest deep expertise in distributed systems, GPU programming, and model optimization.
Arik's outsider status may also be an asset. The AI industry is dominated by a relatively small circle of researchers, investors, and executives who have known each other for years. A fresh voice with a clear commercial thesis and a non-American perspective is exactly what many European investors and customers are looking for. The "sovereign compute" narrative is inherently political, and Arik is fluent in the language of European digital policy in a way that an American founder might not be.
Challenges Ahead
For all the promise, Doubleword faces a steep uphill climb.
Technical complexity is the most immediate challenge. Building an inference optimization platform that works reliably across multiple hardware vendors, model architectures, and deployment environments is extraordinarily difficult. The AI model landscape is fragmented — OpenAI's GPT models, Meta's Llama family, Google's Gemma, Mistral's models, and dozens of others all have different architectures, tokenization schemes, and performance characteristics. Optimizing for all of them is a herculean task. Optimizing for all of them while maintaining compatibility, reliability, and performance is harder still.
Incumbent response is another risk. If Doubleword begins to gain traction, the cloud hyperscalers and API providers it is undercutting will not sit idle. AWS, Google, and Microsoft have armies of engineers and effectively unlimited capital. They can build their own optimization layers, acquire promising startups, or simply lower prices to squeeze out competition. OpenAI, while primarily a model provider, has shown it is willing to compete aggressively on price when threatened.
Customer inertia is a subtler but equally real challenge. Most companies are already deeply integrated with their existing AI infrastructure. Switching to a new platform, even one that promises dramatic cost savings, requires engineering effort, risk assessment, and organizational buy-in. The savings have to be not just real, but obvious and immediate to overcome the friction of change. Doubleword will need to deliver proof points — case studies, benchmark data, and reference customers — to break through.
Funding and scale are perennial startup concerns, but especially acute in AI infrastructure. Building a platform like Doubleword's requires significant capital for hardware, engineering, and customer acquisition. The company has raised funding — the exact amount has not been disclosed — but competing with well-funded rivals like Groq ($1.2 billion raised) and Fireworks AI ($150 million) will require continued access to capital markets. In a funding environment that has tightened since the peak of the AI boom, that is no guarantee.
What This Means for the AI Economy
If Doubleword — or any of the inference optimization startups — succeeds in driving down AI compute costs by an order of magnitude or more, the implications ripple across the entire technology industry.
Cheaper inference means more AI everywhere. Today, AI features are rationed — chatbots are limited to certain hours, image generation is capped, document analysis is restricted to premium tiers. If the cost of running these features drops by 90% or 99%, the economics change completely. Every application becomes an AI application by default. The marginal cost of intelligence approaches zero, just as the marginal cost of storage and bandwidth did in previous technology waves.
Cheaper inference also means more model diversity. Today, most applications use a handful of dominant models because they are the only ones with reliable, affordable API access. If running any model becomes cheap, developers can experiment with smaller, specialized models tailored to specific tasks. A medical diagnosis app might run a fine-tuned healthcare model. A legal document analyzer might use a law-specific model. A game might embed a tiny dialogue model on the player's device. The current oligopoly of foundation model providers gives way to a thriving ecosystem of specialized AI.
And cheaper inference means geographic diversification. If AI can run efficiently on modest hardware, it does not need to be concentrated in massive American data centers. It can run on European clouds, on Asian edge servers, on African mobile devices. The democratization of compute is the democratization of AI capability, and that is a shift with profound implications for global technology competition.
The Bottom Line
Doubleword is a bet on two converging trends: the unsustainable cost of AI inference, and the political imperative for digital sovereignty. At eleven people, it is tiny. At one hundred times cheaper, its claims are audacious. But in an AI economy where the meter is always running, someone has to build the cheaper meter.
Whether Doubleword becomes that someone — or merely a footnote in the history of AI infrastructure — will depend on its ability to deliver on its technical promises, outmaneuver well-funded competitors, and convince skeptical enterprises that the future of AI is not in American clouds, but in infrastructure they control.
For now, it is one of the most interesting early bets in European AI. And in a market that has been starving for homegrown success stories, that is enough to watch it closely.